Why Grade Matters. Bladder cancer grade -- whether a tumor is low-grade or high-grade -- strongly influences treatment decisions. Low-grade non-muscle-invasive tumors are typically managed with transurethral resection alone, while high-grade tumors may require a second surgery or even radical cystectomy. Knowing the grade before operating would allow surgeons to plan more precisely.
Limitations of Current MRI Measures. Diffusion-weighted MRI (DWI) and its derived apparent diffusion coefficient (ADC) maps have shown promise for non-invasive bladder cancer characterization. Prior studies found that high-grade tumors tend to have lower ADC values because densely packed cancer cells restrict water diffusion. However, ADC values alone overlap significantly between low- and high-grade tumors, making accurate grade prediction unreliable with a single number.
Radiomics as a Solution. Radiomics -- the extraction of many quantitative image features beyond simple average signal values -- may capture subtle spatial and textural patterns that better reflect tumor biology. This study tests whether texture features derived from DWI and ADC maps, combined with machine learning, can more accurately grade bladder cancer preoperatively than ADC values alone.
61 Eligible Patients. Seventy-four patients with pathologically confirmed bladder cancer who underwent pretreatment MRI were initially collected. Thirteen were excluded due to very small lesions (under 5 mm) or imaging artifacts that would compromise feature extraction. The final cohort of 61 patients was evenly split: 32 with low-grade and 29 with high-grade tumors as confirmed by surgical pathology.
3.0 Tesla MRI Acquisition. All patients were scanned on the same 3.0 Tesla MRI system using a standardized diffusion-weighted echo-planar sequence with b-values of 0 and 1000 s/mm2. From these two images, ADC maps were calculated voxel by voxel using a monoexponential diffusion model, yielding a spatial map of water diffusion rates across each tumor.
3D Tumor Delineation. Rather than extracting features from a single 2D slice, the researchers outlined the entire tumor in three dimensions by stacking 2D contours drawn on each consecutive MRI slice. Two experienced radiologists independently outlined each tumor and then reached a consensus, blinded to the pathological grade, ensuring consistent and unbiased tumor volumes of interest (VOIs).
102 Features per Tumor. From each 3D tumor VOI on both the DWI and ADC images, two types of texture features were extracted: nine histogram-based features (capturing the statistical distribution of pixel intensity values such as mean, median, and entropy) and 42 gray-level co-occurrence matrix (GLCM) features per image type. In total, 102 features were computed per patient.
What GLCM Features Capture. GLCM features, also called Haralick features, describe the spatial relationships between neighboring pixel intensity values. They quantify properties such as contrast (how much adjacent pixels differ), homogeneity, correlation, and entropy within the tumor. These features are sensitive to the heterogeneity and texture complexity of the tumor tissue, which may differ between low- and high-grade cancers.
Feature Selection Process. Not all 102 features contribute equally to grading. First, a Mann-Whitney U test was applied to all features to identify those with statistically significant differences between low- and high-grade groups (p less than 0.05), reducing the pool to 47 candidate features. A second selection step used support vector machine with recursive feature elimination (SVM-RFE) to rank these 47 features by their contribution to the classification decision and identify the most efficient minimal subset.
Optimal 22-Feature Subset. The SVM-RFE process produced a ranking of features, and the subset containing the top 22 features achieved the highest classification performance as measured by the area under the ROC curve. Among these 22 optimal features, 18 came from GLCM analysis and 4 were histogram features, with 18 derived from ADC maps and only 4 from the raw DWI signal images.
Best Results with the Optimal Subset. The SVM classifier using the 22-feature optimal subset achieved an AUC of 0.861, accuracy of 82.9%, sensitivity of 78.4%, and specificity of 87.1% for distinguishing high- from low-grade bladder cancer. This significantly outperformed using ADC mean alone (AUC 0.675, accuracy 67.2%), demonstrating the added value of textural radiomics.
GLCM Features Outperform Histogram Features. Within the optimal 22-feature subset, the 18 GLCM features achieved an AUC of 0.820 and accuracy of 80.2%, while the 4 histogram features achieved only an AUC of 0.662 and accuracy of 63.7%. This confirms that spatial texture patterns captured by GLCM analysis carry substantially more discriminative power for grading than simple intensity statistics.
ADC Maps Better than Raw DWI. Classification using features derived from ADC maps outperformed features from the raw DWI signal images. The ADC map features achieved an AUC of 0.828 compared to 0.796 for DWI features, suggesting that the quantitative diffusion coefficient values better reflect the biological differences in water mobility between tumor grades than raw signal intensity alone.
Top Three Discriminative Features. The three most important features in the optimal subset were: the GLCM contrast from the ADC map (reflecting local intensity variation between adjacent voxels), the median ADC value (a robust central tendency measure less sensitive to outliers than the mean), and the GLCM entropy from the DWI image (reflecting complexity and randomness in the spatial texture pattern).
GLCM Reflects Tumor Heterogeneity. High-grade bladder cancers are more biologically aggressive, with faster cell proliferation, greater nuclear atypia, and more disorganized tissue architecture. GLCM features capture local spatial heterogeneity in the diffusion signal, quantifying how varied and complex the texture is across the tumor volume, which is expected to be higher in high-grade tumors.
ADC Maps Encode Microstructural Differences. ADC values reflect the restriction of water diffusion by cell membranes. High-grade tumors, which tend to be more densely cellular, restrict water diffusion more strongly, producing lower and more heterogeneous ADC values. GLCM analysis of ADC maps combines both the magnitude and spatial pattern of this diffusion restriction, providing richer information than mean ADC alone.
Signal Intensity Mean Is Not Useful for Grading. The mean signal intensity from the raw DWI image was not significantly different between low- and high-grade tumors (p = 0.583). This makes intuitive sense because raw DWI signal reflects T2 shine-through effects and other non-diffusion factors in addition to true diffusion restriction, making it less specific for tumor grade characterization.
Feature Selection Is Critical. The SVM classifier using all 102 features achieved an AUC of only 0.805 and accuracy of 74.0%, lower than the optimized 22-feature subset. This demonstrates that including too many features, particularly those that are redundant or not informative for grading, actually degrades classifier performance. Systematic feature selection is therefore an essential component of an effective radiomics workflow.
Impact on Surgical Planning. Current clinical practice requires pathological analysis of tumor tissue obtained during transurethral resection (TUR) to determine grade. If high-grade disease is found, guidelines recommend performing a second TUR to ensure complete resection. A reliable preoperative grading method could help identify patients who need more aggressive initial resection or earlier discussion of cystectomy.
FGFR3 Mutations and Low-Grade Disease. Low-grade non-muscle-invasive bladder cancers are closely associated with mutations in the fibroblast growth factor receptor 3 (FGFR3) gene, which has become a therapeutic target. Accurate preoperative identification of likely low-grade tumors could help identify candidates for FGFR3-targeted therapies in the future.
Non-invasive and Reproducible. DWI is already part of routine MRI protocols for bladder cancer evaluation in many centers, meaning that the data needed for this radiomics analysis is available without any additional imaging burden. The approach is non-invasive and could potentially be integrated into the pre-operative workup to complement staging information from T2-weighted MRI.
Small Single-Institution Dataset. With only 61 patients from one institution scanned on the same MRI machine with the same protocol, the generalizability of the trained classifier cannot be fully established. External validation on an independent multi-institutional dataset is essential before clinical translation.
Monoexponential ADC Limitation. This study calculated ADC values using a simple monoexponential diffusion model with just two b-values. More advanced diffusion models such as intravoxel incoherent motion (IVIM) using bi-exponential fitting or stretched exponential models produce additional quantitative parameters (D, D*, and f) that may capture tumor biology more completely and could further improve grading accuracy.
Higher B-Value Potential. Research in prostate cancer has shown that texture features extracted from high-b-value DWI (2000 s/mm2) improve lesion characterization. Future studies could test whether higher b-values similarly improve bladder cancer grading by enhancing contrast between tumors of different grades.
Summary. This study establishes the feasibility and performance of a radiomics approach for preoperative bladder cancer grading using GLCM and histogram texture features from DWI and ADC maps. The proposed workflow, combining systematic feature selection with SVM classification, provides a framework for non-invasive, image-based bladder cancer characterization that warrants validation in larger prospective studies.